Deep Learning for Electromyographic Lower-Limb Motion Signal Classification Using Residual Learning.

Area of Science:

• Biomechanics
• Neuroscience
• Rehabilitation Engineering

Background:

• Electromyographic (EMG) signals are increasingly used for prosthetic and exoskeleton control, primarily in upper limbs.
• Lower limb EMG research, especially with diverse ethnic data and deep learning applications, remains underdeveloped.
• Existing datasets lack standardization and comprehensive muscle coverage for lower limb motion intention.

Purpose of the Study:

• To address the scarcity of lower limb EMG datasets, particularly for Asian populations.
• To introduce the JJ dataset, a large-scale, open-source resource for lower limb EMG analysis.
• To investigate the efficacy of deep learning, specifically ResNet-18, for human motion intention recognition from lower limb EMG signals.

Main Methods:

• Development of the JJ dataset: ~13,350 clean EMG segments from 15 individuals across 10 gait phases, covering nine major leg muscles.
• Signal processing: Utilizing processed time-domain EMG signals as input for deep learning models.
• Classification: Employing an adjusted ResNet-18 architecture for human gait phase recognition.

Main Results:

• The JJ dataset is the first to comprehensively capture nine main muscles involved in human gait.
• Experiments explored preprocessing methods, signal domains (time vs. frequency), and cross-subject recognition accuracy.
• The adjusted ResNet-18 achieved a high average classification accuracy of 95.34% for human gait phases.

Conclusions:

• The JJ dataset provides a valuable, open-source resource for advancing lower limb EMG research.
• Deep learning models like ResNet-18 show significant potential for robust lower limb human motion intention recognition.
• This work highlights the feasibility of using thigh and calf muscle EMG for amputee applications.